human/src/body/decodeMultiple.js

105 lines
5.5 KiB
JavaScript

import * as buildParts from './buildParts';
import * as decodePose from './decodePose';
import * as vectors from './vectors';
function withinNmsRadiusOfCorrespondingPoint(poses, squaredNmsRadius, { x, y }, keypointId) {
return poses.some(({ keypoints }) => {
const correspondingKeypoint = keypoints[keypointId].position;
return vectors.squaredDistance(y, x, correspondingKeypoint.y, correspondingKeypoint.x) <= squaredNmsRadius;
});
}
/* Score the newly proposed object instance without taking into account
* the scores of the parts that overlap with any previously detected
* instance.
*/
function getInstanceScore(existingPoses, squaredNmsRadius, instanceKeypoints) {
const notOverlappedKeypointScores = instanceKeypoints.reduce((result, { position, score }, keypointId) => {
if (!withinNmsRadiusOfCorrespondingPoint(existingPoses, squaredNmsRadius, position, keypointId)) {
result += score;
}
return result;
}, 0.0);
return notOverlappedKeypointScores / instanceKeypoints.length;
}
// A point (y, x) is considered as root part candidate if its score is a
// maximum in a window |y - y'| <= kLocalMaximumRadius, |x - x'| <=
// kLocalMaximumRadius.
const kLocalMaximumRadius = 1;
/**
* Detects multiple poses and finds their parts from part scores and
* displacement vectors. It returns up to `maxDetections` object instance
* detections in decreasing root score order. It works as follows: We first
* create a priority queue with local part score maxima above
* `scoreThreshold`, considering all parts at the same time. Then we
* iteratively pull the top element of the queue (in decreasing score order)
* and treat it as a root candidate for a new object instance. To avoid
* duplicate detections, we reject the root candidate if it is within a disk
* of `nmsRadius` pixels from the corresponding part of a previously detected
* instance, which is a form of part-based non-maximum suppression (NMS). If
* the root candidate passes the NMS check, we start a new object instance
* detection, treating the corresponding part as root and finding the
* positions of the remaining parts by following the displacement vectors
* along the tree-structured part graph. We assign to the newly detected
* instance a score equal to the sum of scores of its parts which have not
* been claimed by a previous instance (i.e., those at least `nmsRadius`
* pixels away from the corresponding part of all previously detected
* instances), divided by the total number of parts `numParts`.
*
* @param heatmapScores 3-D tensor with shape `[height, width, numParts]`.
* The value of heatmapScores[y, x, k]` is the score of placing the `k`-th
* object part at position `(y, x)`.
*
* @param offsets 3-D tensor with shape `[height, width, numParts * 2]`.
* The value of [offsets[y, x, k], offsets[y, x, k + numParts]]` is the
* short range offset vector of the `k`-th object part at heatmap
* position `(y, x)`.
*
* @param displacementsFwd 3-D tensor of shape
* `[height, width, 2 * num_edges]`, where `num_edges = num_parts - 1` is the
* number of edges (parent-child pairs) in the tree. It contains the forward
* displacements between consecutive part from the root towards the leaves.
*
* @param displacementsBwd 3-D tensor of shape
* `[height, width, 2 * num_edges]`, where `num_edges = num_parts - 1` is the
* number of edges (parent-child pairs) in the tree. It contains the backward
* displacements between consecutive part from the root towards the leaves.
*
* @param outputStride The output stride that was used when feed-forwarding
* through the PoseNet model. Must be 32, 16, or 8.
*
* @param maxPoseDetections Maximum number of returned instance detections per
* image.
*
* @param scoreThreshold Only return instance detections that have root part
* score greater or equal to this value. Defaults to 0.5.
*
* @param nmsRadius Non-maximum suppression part distance. It needs to be
* strictly positive. Two parts suppress each other if they are less than
* `nmsRadius` pixels away. Defaults to 20.
*
* @return An array of poses and their scores, each containing keypoints and
* the corresponding keypoint scores.
*/
function decodeMultiplePoses(scoresBuffer, offsetsBuffer, displacementsFwdBuffer, displacementsBwdBuffer, outputStride, maxPoseDetections, scoreThreshold = 0.5, nmsRadius = 20) {
const poses = [];
const queue = buildParts.buildPartWithScoreQueue(scoreThreshold, kLocalMaximumRadius, scoresBuffer);
const squaredNmsRadius = nmsRadius * nmsRadius;
// Generate at most maxDetections object instances per image in
// decreasing root part score order.
while (poses.length < maxPoseDetections && !queue.empty()) {
// The top element in the queue is the next root candidate.
const root = queue.dequeue();
// Part-based non-maximum suppression: We reject a root candidate if it
// is within a disk of `nmsRadius` pixels from the corresponding part of
// a previously detected instance.
const rootImageCoords = vectors.getImageCoords(root.part, outputStride, offsetsBuffer);
if (withinNmsRadiusOfCorrespondingPoint(poses, squaredNmsRadius, rootImageCoords, root.part.id)) continue;
// Start a new detection instance at the position of the root.
const keypoints = decodePose.decodePose(root, scoresBuffer, offsetsBuffer, outputStride, displacementsFwdBuffer, displacementsBwdBuffer);
const score = getInstanceScore(poses, squaredNmsRadius, keypoints);
poses.push({ keypoints, score });
}
return poses;
}
exports.decodeMultiplePoses = decodeMultiplePoses;